TensorFlow Object Detection Accuracy and Loss

Question

I am current using object detection API with my own datasets (5k), and 5 classes (each about 1k).

Model used: ssd_resnet50_v1_fpn_640x640_coco17_tpu-8